singular matrix

[yusofahmad]

In using the program on dummy data to evaluate a model, i got the message that was like "Singluar matrix obtained when estimating outer weights for a block. the problem could be solved by adding more indicators" But the program does not tell me which block this is. Anyways, I had 10 formative LVs and 1 refelctive LV. Maximum formative indictaors on an LV was 10. I had run the procedure on a sample size of 45. I believe a weak rule of thumb argues for 50.

[Diogenes]

Hi Yusof,

1) Firstly I was thinking that your model was with a problem of collinearity (indicators of the formative LV strongly correlated).
Then I had run a model of mine with one more indicator with r=0,99999 with another indicator of the same LV (It was created by simulation). Even in this way the SmartPLS run.

2) In the second run I put a indicator with the same values than another (r=1). This time SmartPLS gave the message that you coted above.


Then, look at your raw data and probably you see that two or more indicators have the same values. Remove one of them of your model and SmartPLS will run (unfortunately your content vality could suffer with this delection, but I couldn´s see a better solution. If you collect a bigger sample will help too, maybe different pattern of responses -between the indicators- will appear...).

I had notice that this problem exist only where the LV is formative.

I hope this help.

Best Regards.

Bido

[jjsailors]

2) In the second run I put a indicator with the same values than another (r=1). This time SmartPLS gave the message that you coted above.


Then, look at your raw data and probably you see that two or more indicators have the same values. Remove one of them of your model and SmartPLS will run (unfortunately your content vality could suffer with this delection, but I couldn´s see a better solution. If you collect a bigger sample will help too, maybe different pattern of responses -between the indicators- will appear...).

I had notice that this problem exist only where the LV is formative.

I hope this help.

Bido

With inner directed indicators PLS uses multiple regression to estimate the weights. What Bido has found would imply that you would encounter the same issue if you created N dummy variables to reflect N levels of a categorical variable (instead of N-1 dummy variables).

I don't think you need worry about content validity, however. Consider that the problem is caused by a variable that is 100% redundant with another variable or group of variables. In this case, dropping it does not cause you to lose any information.

Should not encounter this problem if you specify the indicators as being outer directed.

John

[yusofahmad]

Thanks for the reply, Prof Bido and Sailors. Processing error while using dummy data.. identical coluns used for one formative indicator..problem solved